Dear Gang,

thank you for your help, but I already tried to change my stimuli. It is still not working. Doug offered me another solution that I'd like to share with the whole board.

Hello Emiliano:


Without more details, I can only speculate.


As described on page 20 of the documentation in file 3dDeconvolve.ps, when
the "-stim_nptr k p" command is used, the IRF time limits specified
with "-stim_minlag" and "-stim_maxlag" are in multiples of TR/p. Therefore,
with nptr = 3, and minlag = maxlag = 1, you are modeling the signal as if
it depends on, and ONLY on, the input that occurred TR/3 in the past. This may
result in multicollinearity, depending on the exact nature of the input stim.
functions.


To simplify, suppose there is just one input stimulus function:


myStim.1D: 1 0 0 0 0 0 1 0 0 1 0 0 0 0 0 1 0 0 1 0 0
myData.1D: 7 3 7 7 3 7 7


Save the above (as columns, NOT rows), then execute:


3dDeconvolve \
-input1D myData.1D \
-nfirst 0 -nlast 6 -polort 1 \
-num_stimts 1 \
-stim_file 1 myStim.1D -stim_nptr 1 3 -stim_minlag 1 1 -stim_maxlag 1 1 \
-xout


This yields a non-invertible matrix, since the 3rd column of the X matrix
is all zeros. That is because, in the model definition, only the input at
time lag TR/3 is considered. But the input at time lag TR/3 is always 0.
Note that setting minlag = maxlag = 0 yields an exact fit to the data, for
this particular example.


To verify whether this analysis applies to your experimental design:
Please append the "-xout" command to your script. This will cause 3dDeconvolve
to write the X (design) matrix to the screen. It may be obvious from looking
at the X matrix if there is a problem with multicollinearity (e.g., an entire
column of 0's). If still in doubt about the problem, you could email the
screen output to me.


For more details, see the documentation in Section 1.4.8: Sub-TR Input Stimulus
Functions, in file 3dDeconvolve.ps,


Doug Ward